Reducing the effect of quantization by weight scaling
H.C.A.M. Withagen · 2002
By a statistical analysis of the behaviour of feedforward neural networks to errors in the weights, we show that an optimal scaling factor for the weights exists when the number of inputs to a neuron increases. When this scaling technique is used, the error in the output of a neuron due to quantization errors is not influenced by the size of the network anymore. This technique is especially interesting for the implementation of neural networks using analog electronics.>